It's not "using data registers for mask" that is claimed to be unusual, it's requiring a the mask bit for the 64th byte of a comparison to be usable as as part

It's not "using data registers for mask" that is claimed to be unusual, it's requiring a the mask bit for the 64th byte of a comparison to be usable as as part of the 8th byte of a subsequent addition. I have read that, and it seems like the motivation is that someone who has written a PhD thesis thinks that people don't actually need filtering in their SIMD kernels, which is obviously ridiculous for general-purpose programs manually written in terms of vectors, but which may have been true for the particular programs under study.